Add practical long-audio and memory notes for VibeVoice-ASR

[voidful]

This PR adds practical notes for running VibeVoice-ASR on long recordings and memory-constrained GPUs.

It documents:

• chunked long-audio inference with timestamp-offset stitching;
• coverage, timestamp monotonicity, and repetition checks for long-form ASR;
• cross-chunk speaker-label caveats;
• YaRN RoPE scaling observations for single-pass long-audio robustness;
• memory-oriented Hugging Face generation options such as logits_to_keep=1 and chunked prefill.

Long-Audio Findings

In an 11-item long-form stress test, YaRN with the model's existing context length as the base improved 90-minute robustness:

Setting	e22 90m WER	e22 coverage	TED 90m WER	TED coverage	11-item mean WER	Collapses
No RoPE override	0.5824	77.6%	0.8250	21.8%	0.2328	2
YaRN, factor=1.5, original_max=131072	0.4859	82.0%	0.3422	91.0%	0.2542	0

YaRN is documented as a robustness trade-off, not as a memory optimization.

Testing

Docs-only change; no code tests were run.